Security risk report for @forgecat/anthropics_claude-plugins-official_math-olympiad v0.0.4
Source Integrity
Low
Profile metadata correctly attributes the skill to Anthropic and references the official claude-plugins-official repository.
All source claims (author, original repository, license Apache-2.0, commit hash) are consistent and traceable to a legitimate public source.
No typosquatted, obfuscated, or attacker-controlled package/repository references are present.
Agent Intent
Low
The content describes a legitimate mathematical problem-solving methodology (adversarial verification, pattern-based checking, calibrated abstention) with no instructions to manipulate the AI, ignore system prompts, or exfiltrate data.
Prompts to solver and verifier subagents are pedagogically sound: they instruct reasoning discipline (rigor, honesty about gaps, no tool use for computation) and verification rigor, not deception or hidden behavior.
The 'NO COMPUTATION' instruction in attempt_agent.md is a legitimate constraint to prevent false confidence from brute-force pattern-matching; it does not instruct the AI to hide its instructions or deny its own behavior.
attempt_agent.md: 'NO COMPUTATION. Do not use Bash, Python, WebSearch, Read, Write, or any tool that runs code or fetches data.' — This is a reasoning-discipline constraint, not a manipulation directive.
adversarial_prompts.md: 'Your goal is to find ANY reason to refute.' — Instructs rigorous skepticism, not deception.
SKILL.md: 'Outputs calibrated confidence — will say "no confident solution" rather than bluff.' — Instructs honesty, not role hijacking or credential theft.
Permissions
Low
No tools or hooks are declared in the profile; the skill operates via pure reasoning (subagent prompts) with no file_write, shell, file_delete, or web_fetch authority.
The skill description explicitly states 'Uses pure reasoning (no tools)' and the attempt_agent.md forbids tool use for computation, limiting agency to reasoning only.
No alwaysApply rules, glob patterns, or excessive medium/high-risk categories are present.
Details
Evidence
Profile header: 'Uses pure reasoning (no tools) — then a fresh-context adversarial verifier attacks the proof using specific failure patterns.'
attempt_agent.md: 'You have NO tools — pure reasoning only.' and 'Do not use Bash, Python, WebSearch, Read, Write, or any tool that runs code or fetches data.'
MCP Risk
Low
No MCP servers are declared in the profile.
The skill is self-contained: it uses only subagent orchestration (Claude Code plugin architecture) with no external binary execution, network calls, or filesystem access beyond the reasoning context.
No hidden instructions in tool descriptions or unrestricted access patterns are present.
Details
Evidence
Profile declares '(none)' for MCP servers.
SKILL.md and all referenced files contain only prompt templates and heuristic guidance — no server definitions, binary paths, or network configurations.
ForgeCat